runtime: cleanup M vgetrandom state before dropping P

When an M is destroyed, we put its vgetrandom state back on the shared
list for another M to reuse. This list is simply a slice, so appending
to the slice may allocate. Currently this operation is performed in
mdestroy, after the P is released, meaning allocation is not allowed.

More the cleanup earlier in mdestroy when allocation is still OK.

Also add //go:nowritebarrierrec to mdestroy since it runs without a P,
which would have caught this bug.

Fixes #73141.

Change-Id: I6a6a636c3fbf5c6eec09d07a260e39dbb4d2db12
Reviewed-on: https://go-review.googlesource.com/c/go/+/662455
Reviewed-by: Jason Donenfeld <Jason@zx2c4.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@google.com>

1 files changed, 9 insertions, 2 deletions

diff --git a/src/runtime/vgetrandom_linux.go b/src/runtime/vgetrandom_linux.go
index 00ef2c2d4e..6ab12e3c67 100644
--- a/src/runtime/vgetrandom_linux.go
+++ b/src/runtime/vgetrandom_linux.go
@@ -74,9 +74,16 @@ func vgetrandomGetState() uintptr {
 	return state
 }
 
-func vgetrandomPutState(state uintptr) {
+// Free vgetrandom state from the M (if any) prior to destroying the M.
+//
+// This may allocate, so it must have a P.
+func vgetrandomDestroy(mp *m) {
+	if mp.vgetrandomState == 0 {
+		return
+	}
+
 	lock(&vgetrandomAlloc.statesLock)
-	vgetrandomAlloc.states = append(vgetrandomAlloc.states, state)
+	vgetrandomAlloc.states = append(vgetrandomAlloc.states, mp.vgetrandomState)
 	unlock(&vgetrandomAlloc.statesLock)
 }